Satisfying Information eeds on the Web: a Survey of Web ... · difficulties in the retrieval...

Revista de Estudos Politécnicos

Polytechnical Studies Review

2008, Vol VI, nº 9 ISSN: 1645-9911

Satisfying Information �eeds on the Web: a

Survey of Web Information Retrieval*

Nuno Filipe Escudeiro1•, Alípio Mário Jorge2

•

[email protected], [email protected]

(recebido em 20 de Março de 2008; aceite em 22 de Abril de 2008)

Resumo. Desde muito cedo que a espécie Humana sentiu a necessidade de

manter registos da sua actividade, para que possam ser facilmente consultados

futuramente. A nossa própria evolução depende, em larga medida, deste

processo iterativo em que cada iteração se baseia nestes registos. O

aparecimento da web e o seu sucesso incrementaram significativamente a

disponibilidade da informação que rapidamente se tornou ubíqua. No entanto, a

ausência de controlo editorial origina uma grande heterogeneidade sob vários

aspectos. As técnicas tradicionais em recuperação de informação provam ser

insuficientes para este novo meio. A recuperação de informação na web é a

evolução natural da área de recuperação de informação para o meio web. Neste

artigo apresentamos uma análise retrospectiva e, esperamos, abrangente desta

área do conhecimento Humano.

Palavras-chave: Recuperação de informação na web, motores de pesquisa.

Abstract. Human kind felt, since early ages, the need to keep records of its

achievements that could persist through time and that could be easily retrieved

for later reference. Our own evolution depends largely on this iterative process,

where each iteration is based on these records. The advent of the web and its

* Supported by the POSC/EIA/58367/2004/Site-o-Matic Project (Fundação Ciência e Tecnologia),

FEDER e Programa de Financiamento Plurianual de Unidades de I & D.

1 DEI-ISEP – Deptº de Engenharia Informática, Instituto Superior de Engenharia do Porto ; http://www.dei.isep.ipp.pt

2 2FEP-UP – Faculdade de Economia, Universidade do Porto; http://www.fep.up.pt

• LIAAD, INESC Porto LA – Laboratório de Inteligência Artificial e Análise de Dados; http://www.liaad.up.pt

Tékhne, 2008, Vol VI, nº9

Nuno Filipe Escudeiro, Alípio Mário Jorge

attractiveness highly increased the availability of information which rapidly

becomes ubiquituous. However, the lack of editorial control originates high

heterogeneity in several ways. The traditional information retrieval techniques

face new, challenging problems and prove to be inefficient to deal with web

characteristics. In this paper we present a comprehensive and retrospective

overview of web information retrieval.

Keywords: Web information retrieval, search engines.

1. Introduction

The World Wide Web or simply the web may be seen as a huge collection of

documents freely produced and published by a very large number of people,

without any solid editorial control. This is probably the most democratic – and

anarchic –widespread mean for anyone to express feelings, comments, convictions

and ideas, independently of ethnics, sex, religion or any other characteristic of

human societies. The web constitutes a comprehensive, dynamic, permanently

up-to-date repository of information regarding most of the areas of human

knowledge (Hu, 2002) and supporting an increasingly important part of

commercial, artistic, scientific and personal transactions, which gives rise to a very

strong interest from individuals, as well as from institutions, at a universal scale.

However, the web also exhibits some characteristics that are adverse to the process

of collecting information from it in order to satisfy specific needs: the large volume

of data it contains; its dynamic nature; being mainly constituted by unstructured or

semi-structured data; content and format heterogeneity and irregular data quality are

some of these adverse characteristics.End-users also introduce some additional

difficulties in the retrieval process: information needs are often imprecisely defined,

generating a semantic gap between user needs and their specification.

The satisfaction of a specific information need on the web is supported by search

engines and other tools aimed at helping users gather information from the web.

The user is usually not assisted in the subsequent tasks of organizing, analyzing and

exploring the answers produced. These answers are usually flat lists of large sets of

web pages which demand significant user effort to be explored. Satisfying

information needs on the web is usually seen as an ephemeral one-step process of

Satisfying Information Needs on the Web: a Survey of Web Information Retrieval*

information search (the traditional search engine paradigm). Given these

characteristics, it is highly demanding to satisfy private or institutional information

needs on the web.

The web itself, and the interests it promotes, are growing and changing rapidly, at a

global scale, both as mean of divulgation and dissemination and also as a source of

generic and specialized information. Web users have already realized the potential

of this huge information source and use it for many purposes, mainly in order to

satisfy specific information needs. Simultaneously the web provides a ubiquitous

environment for executing many activities, regardless of place and time.

This paper presents the evolution of information retrieval with particular emphasis

on the web environment. The rest of the paper is organied as follows: section 2

gives a retrospective view of web information retrieval referring to the most

relevant milestones; section 3 defines and describes the web and discusses the

distinction between traditional information retrieval and web information retrieval;

section 4 presents related areas and base theory for web information retrieval;

section 5 discusses research trends in the area and section 6 concludes the paper.

2. A Retrospective View of Web Information Retrieval

Although the need to collect, organize and explore information resources appears

very early in human societies, the advent of the web and digital libraries, together

with advances in communications and computer technology, given the current

challenges, raises the interest on the Information Retrieval field (IR) from scientific

research as well as from commercial and public domains.

Human memory and cognitive capabilities are not able of preserving all knowledge

that human kind produces; we need to store substantial amounts of information

representing this knowledge. These characteristics led human kind to store and

organize information records in a way that facilitates later retrieval and usage for

over 4000 years (Baeza-Yates et al., 1999). Human culture is preserved through

these records.



The common way of referring to information records and allowing for faster

access is through a set of predefined concepts with which are associated related

records. This structure is called the index. Since early IR, indexes represented

taxonomies, hierarchical structures of related concepts. The Dewey Decimal System

is a system of library classification developed in 1876 that has been greatly used

and expanded (Dewey, 2004). These taxonomies, created manually as document

indexing, are reasonable for small document collections but are not scalable.

In 1945, Vannevar Bush claims that scientific efforts should shift from

increasing physical abilities, which had been the main focus during World War II,

to making all previously collected human knowledge more accessible (Bush, 1945).

To help this task he envisions and describes Memex, a system that let users

organize information and relate documents to each other through a mesh of

associative trails running through them. Memex description suggests many

technological features that we use today (hypertext, personal computers, the

Internet, the web, online encyclopaedias such as Wikipedia) and constitutes a main

milestone in the IR area.

The end of World War II released a set of restrictions on scientific and technical

information and generated pressure to make it available to the community. Several

efforts were made to improve the conventional types of index (card catalogs) and

indexing techniques (taxonomies and alphabetical subject headings). In 1950

Mortimer Taube, a government librarian in the USA, realized that a topic list with

40000 subject headings was composed of only 7000 different words. He proposed

Uniterm (Cleverdon et al., 1954), a system that uses these words as index terms and

combines them at search time performing slightly better than the conventional

system, with great reductions of indexing effort (Cleverdon, 1991).

In 1954 the term IR is popularized in a technical report by Cleverdon (Cleverdon et

al., 1954).

During the 60s, IR experienced a period of great evolution, motivated for several

reasons. In 1960, the ex-USSR launches Sputnik I, the first earth artificial satellite,

which generated particular efforts, from the USA, to improve scientific information

spreading; we were living in the Cold War and both opponents struggle for

leadership, especially in the scientific domain. At the same time, the development

of computerized database technology (Codd, 1970) creates conditions for more

efficient IR systems contributing to evolutions in the field. However, although some


developments arise on the automatic categorization of documents (Lim et al., 2001),

indexing still remains mainly manual and only the search becomes mechanical.

It is also during this decade that Cleverdon develops the mathematics for recall

(fraction of relevant documents that are retrieved) and precision (fraction of

retrieved documents that are relevant) and greatly develops the area of IR systems’

evaluation, a main problem in the field. The Cranfield tests apply this technology to

evaluate IR systems on the first test collections compiled with evaluation purpose

(Cleverdon, 1962; Cleverdon et al., 1963).

The hypertext term is coined in 1965 by Ted Nelson who develops Xanadu, a

hypertext system (Nelson, 1965).

In 1968, evidence arises that the value of manual indexing is low and that free-text

indexing may be as effective and much cheaper. In 1971, Jardine and Rijsbergen

establish the cluster hypothesis (Jardine et al., 1971) which states that if a document

satisfies a particular information need, similar documents will probably also satisfy

it. This hypothesis has recently been confirmed for distributed IR (Crestani et al.,

2006).

The vector space model for representing text documents, still the most common

model in IR, is proposed by Salton in 1975 (Salton et al., 1975). Salton is

responsible for the SMART project (Salton et al., 1965), an automatic text

processing system, a standard upon which modern retrieval systems are based.

Three years later, in 1978, takes place the first ACM SIGIR conference.

Rijsbergen published Information Retrieval (Rijsbergen, 1979), a reference book in

IR, with a heavy emphasis on probabilistic models, in 1979. In 1983, Salton

publishes another reference book, Introduction to Modern Information Retrieval

(Salton et al., 1983), focused on vector space models.

During the 70s, IR was viewed as finding the right information in text databases.

In 1989 Tim Berners-Lee proposes a system to manage general information on

research projects at CERN (Berners-Lee, 1989) which became the genesis of the



World Wide Web. The evolution of the web had great impact on our daily lives and

also generated new challenges and a renewed pressure on IR research.

In the early 90s IR moves to full text. Document indexes are built from full content

rather than just abstracts and titles. This decade has been very proficient to IR.

In 1992 the first TREC conference takes place; during this same year the HTML 1.0

formal definition is established. The first web browser, Mosaic which later will

became Netscape, came alive during 1993.

In 1994, MIT and CERN agreed to set up the World Wide Web Consortium (W3C),

an institution aimed at developing and setting standards for the web and web

technologies. This same year, the first full text web search engine, named

WebCrawler, created by Brian Pinkerton at the University of Washington, goes

online and is immediately followed by Lycos. It is the dawning of the first

generation of web search engines that explore mainly on-page data. In 1995

AltaVista comes alive and Yahoo! is incorporated as a company.

With the increase in the number of web users and contents it is realized that the web

is spontaneously becoming a large repository of mankind knowledge and culture.

The Internet Archive project intents to keep records of web content along time to

prevent all this knowledge from being lost (Kahle, 1997). In 2006 this project had

already archived around 2 PB of information, increasing at a rate of 20 TB per

month.

In 1998 two seminal papers, describing the algorithms PageRank (Brin et al., 1998)

and HITS (Kleinberg, 1998), are at the genesis of the second generation of web

search engines. Both these algorithms explore web structure and determine page

quality by analyzing hyperlinks between pages.

The web structure along with web usage patterns are regarded as sources of

evidence that might improve web IR. These are new characteristics that were not

present at traditional IR. In 2000, Broder proposes a model for the web topology

(Broder et al., 2000). In this model, inferred from the mesh of hyperlinks between

web pages, the web is viewed as composed of a central core, made of a set of

strongly interconnected pages; a set of pages with links leading to the central core,


another set of pages pointed to by pages in the central core and a fourth set

composed by relatively isolated pages, that are attached to one of the previous three.

Web page distribution among these four subsets is approximately uniform.

The web lacks editorial control which is one of its benefits but also a drawback. It is

hard to develop systems to automatically process web content due to its

heterogeneity and lack of structure. On the other hand, it is necessary to process

web content automatically, due to its volume and dynamics. In 2001, Berners-Lee,

proposes the semantic web (Berners-Lee et al., 2001). The semantic web is an

attempt to develop a framework that facilitates the automatic processing of web

information.

Lately, web IR systems try to explore multiple sources of evidence aiming to

answer the specific user-need behind the query. Web search engines are now at

their third generation, attempting to merge evidence from various sources, such as:

phrase-based indexing (Hammouda et al., 2004), linguistic approaches (Apostolico

et al., 2006), context (Crestani et al., 2007; Haveliwala, 2005; Ifrim et al., 2005),

temporal analysis (Beitzel, 2006) and semantic models (Siddiqui, 2006).

The evolution of IR systems may be organized in four distinct periods, with

significant differences among the methods that were applied and the sources used

during each one. During an initial period, up to the 50s, the indexing and searching

processes were handled manually. Indexes were based on taxonomies or

alphabetical lists of previously specified concepts. During this phase, IR systems

were mainly used by librarians and scientists.

During a second period, between around 1950 and the advent of web in the early

90s, the pressure on the field and the evolution on computer and database

technology allowed for significant improvements. We went from manual to

automated annotation of documents; however indexes were still built from

restricted descriptions of documents (mainly abstracts and document titles). IR was

viewed as finding the right information in text databases. Operating IR systems

frequently required specific learning. IR systems utilization was expensive and

available only to restricted groups.



During a third period, covering the 90s, the process of indexing and searching

becomes fully automated. Full text indexes are built; web mining evolves and

explores not only content but also structure and usage. IR systems become

unrestricted, cheap, widely available and widely used.

From around 2000 on, the fourth and actual period, other sources of evidence are

explored trying to improve systems’ performance.

Searching and browsing are the two basic IR paradigms on the web (Baeza-Yates et

al., 1999). Three approaches to IR seem to have emerged (Broder et al., 2005):

− the search-centric approach argues that free search has become so good and the

search user-interface so common, that users can satisfy all their needs through

simple queries. Search engines follow this approach;

− the taxonomy navigation approach claims that users have difficulties expressing

their information needs; organizing information on a hierarchical structure

might help finding relevant information. Directory search systems follow this

approach;

− the meta-data centric approach advocates the use of meta-data for narrowing

large sets of results (multi faceted search); third generation search engines are

trying to improve the quality of their answers by merging several sources of

evidence.

IR systems also have to solve problems related to their sources and how to build

their databases/indexes. Several crawling algorithms have been explored, in order to

overcome problems of scale arising from web dimension, such as focused crawling

(Chakrabarti et al., 1999b), intelligent crawling (Aggarwal et al., 2001) and

collaborative crawling (Aggarwal et al., 2004) that explores user behaviour

registered in server logs.

Other approaches have also been proposed: meta-search explores the small overlap

among search engines’ indexes sending the same query to a set of search engines

and merging their answers – a few specific problems arise from this approximation

(Wang et al., 2003); dynamic search engines try do deal with web dynamics, such

search engines do not have any permanent index but instead crawl for their answers

at query time (Hersovici et al., 1998); interactive search (Bruza et al., 2000) wraps


a general purpose search engine into an interface that allows users to navigate

towards their goal through a query-by-navigation process.

At present, IR research seems to be focused on retrieval of high quality, integration

of several sources of evidence and multimedia retrieval.

3. Information Retrieval and Web Information

Retrieval

Classical IR has many applications on the web. However, conventional techniques

must be adapted to the specific characteristics of the web that raise new problems.

Web search engines are the most widely used systems in the web. Recent surveys

(http://searchenginewatch.com, accessed on 2007) estimate that over 75% of web

users explore the web through search services and spend more than 70% of their

time searching online, generating millions of requests daily; Google receives

around 2×109 queries a day. Theses systems are permanently seeking for innovative

ways to achieve the main IR goal: maximize user satisfaction.

3.1 Information Retrieval

IR concerns the study of systems and techniques for representing, organizing and

searching information items. Information items are regarded as unstructured or

semi-structured objects which lead to significant differences from data retrieval

(that concerns structured data) and rises new problems. The IR field is strongly

related to information science, library science and computer science.

3.2 The Web

The web is a public service constituted by a set of applications aimed at extracting

documents from computers accessible in Internet – the Internet is a network of

computer networks. We may also describe the web as an information repository

distributed over millions of computers interconnected through Internet (Baldi et al.,

2003). The W3C defines web in a broad way: “the World Wide Web is the universe

of network-accessible information, an embodiment of human knowledge”.



Due to its comprehensiveness, with contents related to most subjects of human

activity, and global public acceptance, either at a personal or institutional level, the

web is widely explored as an information source. Web dimension and dynamic

nature become serious drawbacks when it comes to retrieving information. Another

relevant characteristic of the web is the absence of any global editorial control over

its content and format. This contributes largely to web success but also contributes

to a high degree of heterogeneity in content, language, structure, correctness and

validity.

Although the problems raised by the size of the web, around 11,5×109 pages (Gulli

et al., 2005), and its dynamics require special treatment, it seems that the major

difficulties concerning the processing of web documents are generated by the lack

of editorial rules and the lack of a common ontology, which would allow for

unambiguous document specification and interpretation. In the absence of such

normative rules, each document has to be treated as unique. In this scenario,

document processing cannot be based on any underlying structure. Although

HTML already involves some structure its use is not mandatory. Therefore, the

higher level of abstraction that may assure compatibility with a generic web

document is the common bag-of-words (Chakrabarti, 2003). This low abstraction

level is not very helpful for automatic processing, requiring significant

computational costs.

The web is a vast and popular repository, containing information related to almost

all human activities and being used to perform an ever growing set of distinct

activities (bank transactions, shopping, chatting, government transactions, weather

report and getting geographic directions, just to name a few). Despite the

difficulties this medium poses to automatic as well as to non-automatic processing,

it has been increasingly explored and has been motivating efforts, from both

academic and industry, which aim to facilitate this exploration.

Currently the web is a repository of documents, the majority of them HTML

documents, that can be automatically presented to users but that do not have a base

model that might be used by computers to acquire semantic information on the

objects being manipulated. The semantic web is a formal attempt from W3C to

transform the web in a huge database that might be easier to process automatically

than our current syntactic web. However, despite many initiatives on the semantic


web (Lu et al., 2002), the web has its own dynamics and web citizens are pushing

the web to the social plan. Collaborative systems, radical trust and participation are

the main characteristics of web2.0, a new paradigm emerging since 2004 (O’Reily,

2004).

3.3 Web Information Retrieval

Web IR is the application of IR to the web. In classical IR, users specify queries, in

some query language, representing their information needs. The system selects the

set of documents in its collection that seem the most relevant to the query and

presents them to the user. Users may then refine their queries to improve the

answer. In the web environment user intents are not static and stable as they usually

are in traditional IR. In the web, the information need is associated with a given

task (Broder, 2002) that is not known in advance and may be quite different from

user to user, even if the query specification is the same. The identification of this

task and the mental process of deriving a query from an information need are

crucial aspects in web IR.

Web IR is related to web mining – the automatic discovery of interesting and

valuable information from the web (Chakrabarti, 2003). It is generally accepted that

web mining is currently being developed towards three main research directions,

related to the type of data they mine: web content mining, web structure mining and

web usage mining (Kosala et al., 2000). Recently another type of data – document

change, page age and information recency – is generating research interests: it is

related to a temporal dimension and allows for analyzing the growth and dynamics

– over time – of the Web (Baeza-Yates, 2003; Cho et al., 2000; Lim et al., 2001).

This categorization is merely conceptual, these areas are not mutually exclusive and

some techniques dedicated to one may use data that is typically associated with

others.

Web content mining concerns the discovery of useful information from web page

content which is available in many different formats (Baeza-Yates, 2003) – textual,

metadata, links, multimedia objects, hidden and dynamic pages and semantic data.

Web structure mining tries to infer knowledge from the link structure on the web

(Chakrabarti et al., 1999a). Web documents typically point at related documents



through a link forming a social network. This network can be represented by a

directed graph where nodes represent documents and arcs represent the links

between them. The analysis of this graph is the main goal of web structure mining

(Donato et al., 2000; Kumar et al., 2000). In this field, two algorithms, which rank

web pages according to their relevance, have received special attention: PageRank

(Brin et al., 1998) and Hyperlink Induced Topic Search, or HITS (Kleinberg, 1998).

Web usage mining tries to explore user behavior on the web by analyzing data

originated from user interaction and automatically recorded in web server logs. The

applications of web usage mining usually intend to learn user profiles or navigation

patterns. Web usage mining is essentially aimed at predicting the next user request

based on the analysis of previous requests. Markov models are very common in

modeling user requests or user paths within a site (Borges, 2000). Association rules

and other standard data mining and OLAP techniques are also explored. (Cooley et

al., 1997) presents an overview of the most relevant work in web usage mining.

Web IR is the evolution of traditional IR applied to the web; however, web

characteristics raise new challenges that are not present in the traditional IR setting.

In web IR, for instance, one has to pay attention to spammers (Chakrabarti, 2003),

people motivated by economic reasons and other, that deliberately add popular

query terms to documents unrelated to those terms (text can be hidden from user if

font and background have the same color, the so called font color spamming); web

documents may be replicated, generating near or exact copies, and complete web

sites may be mirrored in different addresses; web content validity, correctness and

quality are highly variable and difficult to measure; the subject of web documents is

also difficult to determine and the labeling process is expensive and time

consuming. None of these problems arise in traditional IR.

In typical data mining problems, the examples are described in a tabular form where

columns represent the properties, characteristics or features that describe the

concept and each line corresponds to an individual instance or example. Features

are unambiguous; their semantic is precisely known and every single example is

described by exactly the same feature set. In a typical problem from these fields, the

number of features is usually some orders of magnitude smaller than the number of

examples. Learning tasks in web mining do not share many of these desirable


characteristics: the feature space dimension is usually much larger than the number

of available examples (documents), each particular document is described by a

subset of all features and the intersection of these subsets is of reduced dimension,

resulting in a sparse document/term matrix, which creates additional difficulties.

One relevant drawback of a web query, when compared to a database query, which

is due to the lack of structure in web documents, is that it is not usually possible to

answer a web query without some degree of uncertainty (Baeza-Yates, 1999).

While in the database field a query is unambiguous and its mapping to the examples

is precise, in the web environment such precision is not possible since the query

specification – a set of keywords – is usually ambiguous and the relevance of a

certain document to the user information need, specified through that query, is,

therefore, subject to ambiguity and uncertainty. Besides there is a semantic gap

between the real need the user wants to fulfill and the keywords he uses to express

it.

The different aspects between IR and web IR are grouped in four categories by

(Lewandowski, 2005) related to: the documents themselves (including language,

document length, spam, hyperlinks); the user behaviour; the web characteristics

(such as volume, coverage and duplicates) and the search system in itself (user

interface, ranking, search functions).

3.4 Web Search

Web search engines are designed to retrieve the web pages that may answer user

queries. Originally it was implicitly assumed that user intent is to satisfy

information needs; however, recently this assumption was questioned (Broder,

2002). User intention might be informational – the aim is to obtain information on a

given topic – which arises in 40 to 50% of queries, transactional – the aim is to

perform a given transaction, such as shopping or downloading music records –

arising in 30 to 35% of queries, or navigational – the aim is to reach a given site –

arising in 20 to 25% of web queries. The identification of user intent for a given

query may help improving the relevance of the answer and IR systems quality.

As the web evolves, both in volume, heterogeneity and in the set of activities that

users perform on it, search engines utilization increases and the sources of evidence



they explore also evolve. We may classify web search engines according to the set

of features they explore (Broder, 2002). First generation web search engines,

starting in 1994 with WebCrawler and Lycos, explore on-page data (content and

formatting). They support mostly informational queries. The second generation,

emerging in 1998, with Google (PageRank), uses off-page web specific data (link

analysis, anchor text and click streams data) and supports both informational and

navigational queries. The third generation, appearing during the first years of 2000

attempts to merge multiple sources of evidence and aims to support all kinds of

queries.

4. Related Areas and Theory

IR is a multidisciplinary field exploring several scientific areas. Among them we

find library and information sciences – concerned with the collection,

classification, manipulation, storage, preservation, retrieval and dissemination of

information – since early days.

In the late 60s the database technology improved greatly and, from there on has a

major contribution to the field.

IR requires processing non-structured textual data; thus linguistics, natural

language processing and artificial intelligence also contribute greatly to the field.

The advent of the web turns information into a ubiquitous and heterogeneous

resource and increases the volume of publicly available information to huge

dimensions. In this setting the statistical and machine learning fields become

fundamental to IR.

The web itself and the way people build and use its content are social phenomena in

nature, and exploring it may benefit from social network analysis, which is another

relevant area contributing to the IR field.

Merging contributions from all these and many other scientific areas generates

solutions for many problems that arise in the IR field, such as: the categorization

and clustering of documents, relevance evaluation, cross-language retrieval,


distributed IR, question answering, spam detection, collaborative filtering, content

management and adaptive web sites, recommendation, results presentation.

Web IR requires the automation of web page processing; otherwise it is not scalable

for the web dimension and dynamics.

For the sake of IR we may group web page processing issues in three stages:

− Pre-processing: converting a web page into a manageable representation;

includes a data preparation phase (removes irrelevant data and adds metadata)

and a data representation phase (maps these data structures to adequate models);

− Learning: inferring knowledge from the available data with the aim of finding

and organizing web pages in a way that satisfies user needs;

− Presentation: presenting relevant pages to the user in a way that allows for an

efficient exploration. Although this is not usually a concern of web search

engines, it seems that using visual tools to organize and present large document

collections might improve user satisfaction.

4.1 Pre-processing

The pre-processing stage concerns the tasks that are required to obtain a suitable

web document representation, valid for the subsequent automatic learning phase.

This stage includes data preparation and data representation.

The data preparation phase includes several steps that attempt to eliminate non-

informative features. This phase usually includes (Baeza-Yates et al., 1999):

lexical analysis – eliminating punctuation, accents, extra spacing, converting to

lower/upper case;

stop-word removal – removing irrelevant terms; requires a list of words to eliminate

(stop-words);

stemming – reducing words to their semantic root; Porter algorithm (Porter, 1980) is

probably the most well known stemming algorithm, there are specific algorithms

for the Portuguese language (Orengo et al., 2001);

indexing – defining index terms, the features that will be used for document

modeling, and computing their weights.



The application of these pre-processing tasks must be careful because the predictive

power of words is highly dependent on the topic of interest (Chakrabarti et al.,

1998a). Another essential aspect to consider is the language in which the document

is written, which determines the list of stop-words and the stemming algorithm to

use. Besides, stop-word removal always reduces information contained in the

document; to avoid this loss some search services, like CiteSeer (Lawrence et al.,

1999), do not remove any words from the documents to be indexed. Web

documents written in HTML still require removing HTML tags.

In the web environment some other problems have to be dealt with, as, for instance,

document duplicates that may exist and which should be detected by the system.

While exact match between two documents is easy to detect the near duplicates

problem is harder to deal with. Duplicate web pages may have slight differences

between them – a date, the home URL or editor, to name a few, which raise

difficulties to automated duplicate detection.

It is possible to estimate with a reasonable confidence the characteristic distribution

of a set of features when the number of training examples is substantially higher

than the number of distinct features, which is not the case in web page

classification. In text mining the number of features is typically much larger than

the number of available training examples and, if care is not taken undesirable

overfitting may arise. Feature selection is desirable not only to avoid overfitting but

also to keep the same level of accuracy while reducing feature space dimension, and

consequently reducing the necessary computational effort. Feature reduction or

feature selection techniques may be heuristic – governed by linguistic principles or

specific rules from the universe of discourse – or statistical.

Feature selection techniques, in text classification, are usually applied following the

process (Chakrabarti, 2003):

1. compute, for each feature, a measure that allows to discriminate categories;

2. list features in decreasing order of that measure and

3. keep the subset of the features with the highest discriminative power.

The curse of dimensionality (Koller et al., 1996) requires reducing the feature space

dimensionality in order to improve any reasoning over this space. The high feature


space dimensionality can be reduced with techniques that might be categorized as

feature selection or re-parameterisation techniques (Aas et al., 1999).

Feature selection attempts to remove non-informative words from documents in

order to improve categorization effectiveness and reduce computational complexity

while re-parameterisation is the process of constructing new features, as

combinations or transformations of the original ones. (Yang et al., 1997) describes

common feature selection techniques.

It should be stressed that words in the document are not the only features in web

pages. Hypertext documents contain other types of features, such as links to and

from other web pages and HTML tags, which may be explored, in isolation or in

conjunction with others, constituting abstract features (Halkidi et al., 2003) which

may hold significant predictive power.

After the data preparation phase, each document is reduced to its representative

features and the next step, data representation, is to encode this view of the

document into a specific format, ready for the learning stage.

Classic IR models (Boolean, vector and probabilistic) assume that each document is

described by a set of representative keywords called index terms (Baeza-Yates et

al., 1999). Index terms are words, or phrases, appearing in the document, whose

semantics helps in remembering the document’s main themes. Index terms are

usually assumed to be independent.

From the Boolean model perspective, a query is a Boolean expression using the

three connectives “and”, “or” and “not”. The weights associated with index terms

are binary. There is no notion of partial match; each document is either relevant or

not relevant at all to a given query; the lack of a relevance notion makes this a data

retrieval instead of a real IR model. Simplicity and the clean formalism are the main

advantages of this model while its main disadvantages come from the fact that exact

matching may lead to retrieval of too many – poor precision – or too few – poor

recall – documents.



The Vector model, probably the most commonly used, assigns real non-negative

weights to index terms in documents and queries. In this model, documents are

represented by vectors in a multi-dimensional Euclidean space. Each dimension in

this space corresponds to a relevant term/word contained in the document

collection.

The degree of similarity of documents with regard to queries is evaluated as the

correlation between the vectors representing the document and the query which can

be, and usually is, quantified by the cosine of the angle between the two vectors.

In the vector model, index term weights are usually obtained as a function of two

factors:

− the term frequency factor, TF, a measure of intra-cluster similarity; computed as

the number of times that the term occurs in document, normalized in a way as to

make it independent of document length and

− an inverse document frequency, IDF, a measure of inter-cluster dissimilarity;

weights each term according to its discriminative power in the entire collection.

This model main advantages are related to improvements in retrieval performance

due to term weighting; partial matching that allows retrieval of documents that

approximate the query conditions. The index term independency assumption is

probably its main disadvantage.

Some proposals, distinct from the traditional vector space model, try to explore

sequences of characters, using kernel functions to measure similarity between

documents; the string kernel model. Text can further be represented as sequences of

words, which are linguistically more meaningful than characters (Nicola et al.,

2003).

Probabilistic models compute the similarity between documents and queries as the

odds of a document being relevant to a query. Index term weights are binary. This

model ranks documents in decreasing order of their probability of being relevant,

which is an advantage. Its main disadvantages are: the need to guess the initial

separation of documents into relevant and non-relevant; weights are binary; index

terms are assumed to be independent.


Classic models for text retrieval view a document in its most primitive form: a text

is a bag-of-words; the eventual text structure is not explored. Retrieval models,

which combine information on text content with information on the document

structure, are called structured text retrieval models. Web pages, which are mainly

hypertext documents contain a set of features, not found in plain text documents,

which can reveal to be highly informative. Hypertext features on web pages

include: HTML tags, URLs, IP addresses, server names contained in URL,

sub-strings contained in URLs, links from the current page to other pages – out-

links – and links from other pages to the current page – in-links.

At the end of the 1980s and throughout the 1990s, various structured text retrieval

models have appeared in the literature (Charkrabarti, 2003): non-overlapping list

model, proximal nodes model, simple concordance list models (Dao, 1998), path

prefixing, inductive classifiers, such as FOIL, applied to relational models.

4.2 Learning

Automated learning techniques, generally adopted from the machine learning and

statistics fields, may improve the performance and the functionality of IR systems

in a wide variety of forms, ranging from document classification to user behaviour

modeling and information extraction.

Document classification is the task of assigning one or more pre-defined categories

to documents. In the web environment we are usually interested in sets of several

distinct classes; this is known as a multi-class problem. Besides, web documents are

often related to more than one category, the multi-label characteristic. The problem

of classification in IR is a multi-class and multi-label problem.

Text classifiers, from text mining field, deal primarily with flat text documents and

do not take advantage of other potentially relevant features present in web

documents. Web pages contain other features, besides flat text, like hyperlinks,

content of neighbours and metadata that might help to improve classifiers’

performance.

Several text mining algorithms, derived from the machine learning field, have been

applied to the web document classification task. Although the performance of the

text classifiers depends heavily on the document collection at hand (Yang, 1999),



some classifiers, particularly Support Vector Machines (SVM) and

k-Nearest-Neighbours seem to outperform others in the majority of the domains.

(Joachims, 1998) points out a few properties of text – high dimensional feature

spaces, few irrelevant features, document vectors are sparse, most text

categorization problems are linearly separable – that justify why SVM should

perform well for text categorization.

When applied to web pages, classical flat text classifiers treat each page

independently, ignoring links between pages and the class distribution of neighbour

pages.

Some algorithms explore other sources of evidence, besides plain text. (Yang et al.,

2002), define five types of regularities that might be present in a hypertext

collection: no regularity – the only place that has relevant information about the

class of the document is the document itself, encyclopaedia regularity – documents

with a given class only link to documents with the same class, co-referencing

regularity – some, or all, of the neighbouring documents belong to the same class

but this class is distinct from the document class, pre-classified regularity – the

regularity is present at the structural level where a single page (hub) points to

several pages which belong to the same class – and metadata regularity – metadata

is available from external sources and can be explored as additional sources of

evidence generating new features. The authors then define the types of features that

should be used in order to improve the classification task of documents belonging

to each of these regularities.

(Chakrabarti et al., 1998b) also test several feature sets, similar to the ones

suggested by (Yang et al., 2002): local text, local text concatenated with all

neighbours’ text, local text plus neighbours text prefixed with discriminative tags,

to conclude that naive use of terms in the link neighbourhood of a document can

even degrade performance.

In document classification, the class labels are frequently organized in taxonomies

where classes are associated through inheritance. Techniques that explore this


structure are commonly referred to by hierarchical classification (Chakrabarti et al.,

1998a).

Evaluation of systems’ performance is a very important issue in IR; TREC is an

annual conference especially devoted to this issue. Conventional evaluation is based

on test collections. Test collections are composed by a set of documents, a set of

queries and a set of relevant documents for each query or the specification of the

relevance criteria to apply (Cormack et al., 1998; Rijsbergen, 1979 – chapter 7;

Voorhees, 1998).

This evaluation can be based on several measures. Among them recall – ratio

between the number of documents correctly classified and the total number of

documents in the category – and precision – ratio between the number of

documents correctly classified and the total number of documents classified in the

category – assume a very important position.

Usually a classifier exhibits a trade-off between precision and recall, so these

measures are negatively correlated: improvement in recall is made at the cost of

precision and vice-versa. It is frequent to have text classifiers operating at the

break-even point. The break-even point of a classification system is the operating

point where recall and precision have the same value.

Accuracy – ratio of the number of correctly classified examples by the total number

of evaluated examples – and error – ratio of the number of incorrectly classified

examples by the total number of evaluated examples – are complementary measures

of the error probability of the classifier given a certain category.

These measures are defined for binary classifiers. To measure performance in

multi-class problems there are two common approaches to aggregate these

measures evaluated at each singular category: macro-averaging and

micro-averaging. Macro-averaged measures are obtained by first computing the

individual scores, for each individual category, and then, averaging these scores to

obtain the global measure. Micro-averaging measures are obtained by first

computing the total number of documents correctly and incorrectly classified, for



all of the categories, and then using these values to compute the global performance

measure, applying its definition. Macro-averaging gives equal weight to each

category while micro-averaging gives equal weight to every document.

In current web IR, it is typical that users specify their information need with a set of

keywords. Although this form of interaction is very simple and relies on universally

understood specification language and protocol, keywords are weak primitives to

specify an information need (with high imprecision and subject to each user’s

interpretation), resulting in ambiguous, incomplete or excessive specifications and,

consequently, poor result sets. These result sets are analysed and validated by the

user. The user implicitly analyses the retrieval system performance when he

explores the result set. It is reasonable to expect that if a user downloads many of

the returned documents, then they probably have some relevance; and, on the other

hand, if the user does not download a given document it is probable that this

particular document is not very relevant to the user need. The user may even be

required to explicitly indicate perceived relevance for a given set of documents.

This feedback from the user, either explicit or implicit, might be used in order to

improve its performance, particularly as to the relevance of returned documents;

this problem is known as relevance feedback. Relevance feedback techniques

usually produce some query modifications, by adding, removing or changing terms,

originating internal queries that are different from the original user query and which

are expected to be more representative of user’s needs. The common approach is to

retrieve an initial set of relevant (and eventually also non-relevant) documents and

discover correlated features that can be used to narrow down the original query,

improving precision (Chakrabarti, 2003; Glover et al., 2001; Mitra et al., 1997).

4.3 Presentation

In IR systems the user typically interacts with the system in two distinct moments:

when specifying information needs and when analysing the results.

Query specification requires the user to describe information needs as a set of

keywords and, eventually, some other characteristics to be met, such as the domain

and file format. In traditional search engines the user specifies the query in one


single step, while interactive query systems (Bruza et al., 2000) require the user

interaction during the initial tuning phase where the user iteratively refines the

query by following suggestions from the system.

The analysis of the results is a valuable aspect of search services, however, public

search services have not been given it much of importance. Results from a search

engine are usually presented as a flat ranked list. Retrieval systems are a black box

from the user point of view – the user submits a query and receives an answer

without having any knowledge on the decisions, processes and data analysed to

provide the answer – which obstructs the analysis of the results. Besides, in a flat

list, users do not have a way of analysing a global perspective of the entire

collection neither of analysing a single object while still keeping a global image of

all the collection. Several difficulties may arise from these facts (Cugini et al.,

1996; Olsen et al., 1992).

A visualization tool might help in solving some of these difficulties, providing

functionalities that allow for selecting valuable information from large document

collections without much effort. The flat list approach is very restrictive when the

object collection is large (Carey et al., 2000); this approach subordinates the user to

a slow, intensive mental process of reading through the reference list while in a

graphical system the user just has to recognize patterns on a visual display, which is

a much faster and less demanding process.

Post-retrieval visualization techniques may be categorized, according to their main

goal, in the following categories (Zamir et al., 1999): on one hand those that stress

inter-document similarities, usually based on document content, and which help the

user getting an overview of the full document collection as well as visualizing

clusters of topically related documents and, on the other hand, those that aim to

display additional information about retrieved documents, which might be

predefined document attributes (such as size, date, age, author or source) or user-

specified attributes (such as predefined categories).

Although public search services rarely explore post-retrieval visualization

techniques, there are many distinct proposals. We briefly review some of them:



(Carey et al., 2000) claim that an unified approach – allowing the user to explore

the corpus through several different visualization paradigms, which should be

integrated, in order to allow for cross comparisons and evaluations – will be

valuable. They include Sammon Maps, Tree-Map visualization and Radial

visualization in this consolidated framework.

Sammon maps try to represent n-dimensional objects in a 2-dimensional space

while attempting to preserve the pair-wise distances between objects.

Tree-Map visualization represents clusters as rectangles arranged in such a way as

to fit inside a larger rectangle. The area of a rectangle is proportional to the size of

the cluster it represents. Similar clusters are grouped together in super-clusters.

Radial visualization is a common paradigm where keywords are represented as

nodes, which are homogenously displayed, usually around a circumference, in the

display space. Documents are then placed as if they were connected to the relevant

keyword nodes by forces that attract them proportionally to the weight of the word

in the document.

The VIBE system (Olsen et al., 1992) explores this radial visualization paradigm by

defining Points Of Interest as a set of representative keywords working as anchor

nodes relatively to which documents are placed in a bi-dimensional space.

(Spangler et al., 2003) represents relevant categories, associated to the query result

set, in a Radial graph, allowing the user to understand which categories are strongly

related to the query and associates this view with a binary tree that allows to refine

a query, once the user as chosen a given category.

(Viji, 2002) describes a visualization tool that uses springs to represent documents

semantic correlation. The length of a spring connecting two documents is

proportional to their correlation. Correlation between two documents is based on

textual content and link based information.

The Document Spiral paradigm, proposed in (Cugini et al., 1996) displays

documents along a spiral in a two dimensional space. Documents at the centre of

the spiral are the most relevant; relevance decreases as one travels along the spiral.


Kohonen maps apply neural network principles to self-organize document

collections into a galaxy – a galaxy is simply a group of clusters represented in a

two-dimensional space, one of the simplest forms of cluster mapping (Chewar et

al., 2001).

Star coordinates is an exploratory visualization technique for organizing and

representing n-dimensional spaces on a two-dimensional surface. Together with

interactive functionalities is an effective tool to explore large corpus and to detect

clusters (Kandogan, 2001).

5. Research in Information Retrieval

The amount, heterogeneity and dynamics of available information on the web along

with the vast number of activities that are available online demand for new

approaches to the problem of searching the web. Understanding the task that users

are trying to perform becomes a relevant part of the problem together with the

perception of their specific need. Search systems still struggle to deliver quality

answers and new sources of evidence are being explored.

Open research problems on the IR field are recognized by several experts. We

resume some of these perspectives.

(Baeza-Yates et al., 1999) identified a set of research directions, organized by

specific areas on the IR field. Among them we selected those that seem more

relevant:

− Retrieval of higher quality;

− Combining several evidential sources to improve relevance judgements;

− Querying on the structure of the text besides content. Visual query languages;

− Interactive user interfaces;

− New retrieval strategies, based on the understanding of user behaviour;

− User-centered design; cognitive and behavioural issues;

− Understanding the criteria by which users determine if retrieved information

meets their information needs.



This list stresses the importance of improving retrieval quality, combining several

evidential sources and focusing the retrieval process on the user.

(Chakrabarti, 2003) refers, among other, the importance of combining several

external sources of evidence and personalization of search systems:

− Disassembling pages and page zones into finer structures, such as phrases and

sentences;

− Integration of several external sources of evidence, such as lexical networks

(WordNet), thesaurus and ontologies;

− Profiles, personalization and collaborative search;

− Modeling of semantics (web pages, user, user needs).

(Croft, 2003) refers to global information access and contextual retrieval:

− Leveraging worldwide structured and unstructured data in any language;

− Combine search technologies, knowledge about query and user context to

provide the most appropriate answer for a user’s information need.

(Henzinger et al., 2003) refers a more specific list of problems:

− Spam; pro-active approaches to detect and avoid spam (text spam, link spam,

cloaking or combined), collectively known as adversarial search;

− Combine link-analysis quality judgments with text-based judgments to improve

answer quality;

− Estimating web page quality from web structure (within a page, across different

pages);

− Quality evaluation; implicit relevance feedback, click-through data;

− Detect and explore web conventions; understanding the nature of links

(commercial, editorial, metadata);

− Vaguely structured data; try to infer semantic information from HTML tags,

since layout conveys semantic information.


(Shwarzkopf, 2003) emphasizes the importance of conveniently organizing

documents in the answer. Interaction with information does not end with retrieving

relevant documents; it also includes making sense of the retrieved information,

organizing collected materials according to user needs.

(Sahami, 2004) also refers to high quality search results, dealing with spam and

search evaluation:

− Identify which pages are of high quality and relevance to a user’s query;

− Linked-based methods for ranking web pages;

− Adversarial classification, detecting spam;

− Evaluating the efficacy of web search engines;

− Determining the relatedness of fragments of text, web contextual kernel;

− Retrieval of images and sounds;

− Harnessing vast quantities of data.

(Apostolico et al., 2006) stress the importance of query expansion, search

evaluation and retrieval from XML sources:

− Measures to assess IR system’s efficiency and to compare it with others;

− Efficient query expansion;

− Query performance prediction, particularly in the case of query expansion;

− Retrieval model and query language for XML documents.

Among all these perspectives on IR research we detect some common trends:

Retrieval of high quality seems to be still the most relevant aspect to be solved. The

majority of research efforts on IR follow this major goal. Recently, as the amount of

human activity online increases, the task that the user is trying to perform while

using IR systems assumes relevance as an indicator of what is the real need behind

the query (Broder, 2002). Understanding and classifying user queries is an

important step (Betzel, 2006). Information overload is a web characteristic that

requires high quality retrieval; otherwise IR systems will fail their goal because

they will not be able to produce answers made of (small) sets of relevant



documents. Besides, this high quality must be achieved without requiring explicit

user effort. Semi-supervised classification methods (Li et al., 2003; Nigam et al.,

2000; Bennet et al., 1998; Blum et al., 1998), specifically applied on the web

environment might improve retrieval quality while reducing user’s workload.

Conditional Random Fields (Lafferty et al., 2001) may also help improving

retrieval quality.

Web IR is also being applied to new specific ways of using the web. TREC

introduced a new track in 2006 that deals with retrieval from the blogosphere.

(Sahami, 2004) refers to specific methods for ranking UseNet or bulletin board

postings. The Web 2.0 paradigm (O’Reilly, 2004), based on active users, reinforces

the dynamic nature of the web and originates new challenges and opportunities.

Integration of several sources of evidence is being explored by researchers trying to

improve the modeling of users and user needs. Several distinct features are being

considered and analyzed: linguistic approaches (Arcot, 2004), context sensitive

search (Crestani, 2007; Haveliwala, 2005; Ifrim et a, 2005; Zakos et al., 2006), task

or topic-based analysis of queries (Beitzel, 2006), query-dependent PageRank

(Richardson et al., 2004), Wikipedia-assisted feedback (Liu et al., 2005), semantic

models (Siddiqui et al., 2006) and phrase-based indexing (Hammouda et al., 2004)

are some examples of research on this subject. Exploring new ways of integrating

distinct feature sets, such as Formal Concept Analysis (Shen, 2005; Wolf, 1993) or

Markov Logic (Domingos, 2007: Domingos et al., 2006) may produce interesting

results.

Personalization issues are also being explored (Escudeiro et al., 2006). Information

needs are user specific and IR systems should provide user specific answers,

organized and presented according to particular users or groups of users’ specific

interests.

Besides these long breath problems there are a few more specific problems

generating research interest, such as adversarial search that deals with spam,

retrieval from XML sources and exploring web conventions.


6. Conclusions

The web is a vast repository of information with some characteristics that are

adverse to IR: large volume of data, mainly unstructured or semi-structured;

dynamic nature; content and format heterogeneity and irregular data quality are

some of these adverse characteristics. These specific web characteristics require

specific treatment.

The user also introduces additional difficulties to the retrieval process, such as the

semantic gap, arising from ambiguous query specifications and the fact that the

required organization of the answer and the aim the user is seeking are not fed to

the IR system.

Despite these difficulties the web is being used as an information source as well as

a support for an increasing number of activities by an increasing number of people

with rather distinct background, motivations and needs.

All these aspects give us reasons to believe that the web IR field is, and will remain,

relevant and challenging to academic and economic areas.

References

Aas, K., Eikvil, L. (1999), Text Categorization: A Survey, Norwegian Computing Center Aggarwal, C.C., Al-Garawi, F., Yu, P. (2001), Intelligent crawling on the World Wide Web with

arbitrary predicates, Proceedings of the 10th World Wide Web Conference

Aggarwal, C.C. (2004), On Leveraging User access Patterns for Topic Specific Crawling, Data mining and Knowledge Discovery, 9, pp 123-145, Kluwer Academic Publishers

Apostolico, A., Baeza-Yates, R., Melucci, M. (2006), Advances in information retrieval: an introduction

to the special issue, Journal of Information Systems, Elsevier Science Ltd., 31(7), p.569-572 Arcot, H.G.A. (2004) Perception-based fuzzy information retrieval. United States -- California: San Jose

State University

Baeza-Yates, R. (2003), Information Retrieval in the Web: beyond current search engines, Elsevier International Journal of Approximate Reasoning, 34, pp 97-104

Baeza-Yates, R., Ribeiro-Neto, B. (1999), Modern Information Retrieval. ACM Press

Baldi, P., Frasconi, P., Smyth, P. (2003), Modeling the Internet and the Web. Probabilistic Methods and Algorithms, Wiley

Beitzel, Steven M. (2006) On understanding and classifying web queries, PhD dissertation USA, Illinois, Illinois Institute of Technology

Bennet, K.P., Demiriz, A. (1998), Semi-Supervised Support Vector Machines, Proceeding of 3eural

Information Processing Systems Berners-Lee, T. (1989), Information Management: a proposal., CERN



Berners-Lee, T., Hendler, J., Lassila, O. (2001), The Semantic Web. Scientific American Blum, A., Mitchell, T. (1998), Combining labelled and unlabelled data with Co-training, Proceedings of

the 11th Annual Conference on Computational Learning Theory, pp 92-100

Borges, J.L.C.M. (2000), A Data Mining Model to Capture User Web 3avigation Patterns, PhD dissertation, University of London

Brin, S., Page, L. (1998), “The anatomy of a large-scale hypertextual web search engine”, Proceedings

of the 7th World Wide Web Conference, pp 107-117 Broder, A. (2002) A taxonomy of web search. SIGIR Forum. 36:2. p. 3-10

Broder, A., Kumar, R., Maghoul, F., Raghavan, P., Rajagopalan, S., Stata, R., Tomkins, A., Wiener, J.

(2000), Graph structure in the web World Wide Web Conference, Amsterdam, Holand Broder, A., Maarek, Y., Bharat, K., Dumais, S., Papa, S., Pedersen, J., Raghavan, P.(2005), Current

Trends in the Integration of Searching and Browsing, Special interest tracks and posters of the 14th

World Wide Web Conference , Chiba, Japan, p.793 Bruza, P., McArthur, R., Dennis, S. (2000), Interactive Internet search: keyword, directory and query

reformulation mechanisms compared, Research and Development in Information Retrieval

Bush, V. (1945), As We May Think, The Atlantic Monthly, July Carey, M., Kriwaczek, F., Ruger, S.M. (2000), A Visualization Interface for Document Searching and

Browsing, Proceedings of the 3PIVM 2000

Chakrabarti, S. (2003), Mining the Web. Discovering Knowledge from Hypertext Data. Morgan Kaufmann Publishers

Chakrabarti, S., Dom, B., Agrawal, R., Raghavan, P. (1998a), Scalable feature selection, classification

and signature generation for organizing large text databases into hierarchical topic taxonomies, The VLDB Journal, 7, pp 163-178

Chakrabarti, S., Dom, B., Indyk, P. (1998b), Enhanced hypertext categorization using hyperlinks,

Proceedings of ACM SIGMOD International Conference on Management of data, pp 307-318 Chakrabarti, S., Byron, E., Kumar, S., Raghavan, P., Rajagopalan, S., Tomkins, A., Gibson, D.,

Kleinberg, J. (1999a), Mining Web Link Structure, IEEE Computer, 32(8), pp 60-67

Chakrabarti, S., Berg, M., Dom, B. (1999b), Focused crawling: a new approach to topic-specific resource discovery, Proceedings of the 8th World Wide Web Conference

Chewar, C.M., Krowne, A., O´Laughlen, M. (2001), User Object Collections: Visualization Concepts by

collection-Insight 3eed, CITIDEL project Cho, J., Garcia-Molina, H. (2000), Estimating Frequency of Change, Technical report, Stanford

University

Cleverdon, C.W. (1991), The significance of the Cranfield tests on index languages, Proceedings of the ACM – SIGIR, p. 3-12

Cleverdon, C.W. (1962), Comparative Efficiency of Indexing Systems, Cranfield

Cleverdon, C.W., Aitchison, J. (1963), Test of the Index of Metallurgical Literature, Cranfield Cleverdon, C.W., Thorne, R.G. (1954), An Experiment with the Uniterm System, R.A.E. Cranfield, 7

Codd, E.F. (1970), A Relational Model of Data for Large Shared Data Banks, Communications of the

ACM, Vol. 13, No. 6, June 1970, pp. 377-387 Cooley, R., Mobasher, B., Srivastava, J.(1997), Web Mining: Information and Pattern Discovery on the

World Wide Web, Proceedings of the 9th IEEE International conference on tools with Artificial

Intelligence, pp 558-567 Cormack, G.V., Palmer, C.R, Clarke, C.L.A. (1998), Efficient Construction of Large Test Collections,

Proceedings of the ACM SIGIR 1998 Conference

Crestani, F., Shengli, W. (2006), Testing the cluster hypothesis in distributed information retrieval, Information Processing and Management. 42, p. 1137-1150

Crestani, F., Ruthven (2007), I., Introduction to special issue on contextual information retrieval systems.

Information Retrieval. 10, p. 111-113


Croft, W.B. (2003), Information retrieval and computer science: an evolving relationship, ACM SIGIR Conference, Toronto, Canada, p.2-3

Cugini, J., Piatko, C., Laskowski, S. (1996), Interactive 3D Visualization for Document Retrieval,

Proceedings of the ACM Conference on Information and Knowledge Management Dao, T. (1998), An Indexing Model for Structured Documents to Support Queries on Content, Structure

and Attributes, Proceedings of IEEE ADL Conference, Santa Barbara, California, USA

Dewey, M. (2004), A Classification and Subject Index for Cataloguing and Arranging the Books and Pamphlets of a Library, Project Gutenberg EBook

Domingos, P. (2007), What's missing in AI: The Interface Layer, University of Washington,

Washington, USA Domingos, P., Kok, S., Poon, H., Richardson, M., Singla, P. (2006), Unifying Logical and Statistical AI,

The Twenty-First National Conference on Artificial Intelligence and the Eighteenth Innovative

Applications of Artificial Intelligence Conference, Boston, Massachusetts, USA Donato, D., Laura, L., Millozi, S. (2000), A beginner’s guide to the Webgraph: Properties, Models and

Algorithms, Proceedings of the 41st FOCS, pp.57-65

Escudeiro, N., Jorge, A., (2006) Semi-automatic Creation and Maintenance of Web Resources with webTopic. Semantics, Web and Mining. LNCS, vol. 4289, pp. 82-102, Springer, Heidelberg

Glover, E.J., Flake, G.W., Lawrence, S., Birmingham, P., Kruger, A., Giles, C.L., Pennock, D.M.

(2001), Improving Category Specific Web Search by Learning Query Modifications, Symposium on Applications and the Internet, IEEE Computer Society, pp 23-31

Gulli, A., Signorini A. (2005), The Indexable Web is More than 11.5 billion pages. In: WWW 2005,

Chiba, Japan Halkidi, M., Nguyen, B., Varlamis, I., Vazirgiannis, M. (2003), “Thesus: Organizing Web document

collections based on link semantics”, The VLDB Journal, 12, pp 320-332

Hammouda, K.M., Kamel, M.S. (2004), Efficient Phrase-Based Document Indexing for Web Document Indexing. IEEE Transactions on Knowledge and Data Engineering. 16:10, p. 1279-1296

Haveliwala, T.H. (2005), Context-sensitive Web search, PhD dissertation, Stanford University,

California, USA Henzinger, M., Motwani, R., Silverstein, C. (2003), Challenges in Web Search Engines, 18th

International Joint Conference on Artificial Intelligence

Hersovici, M., Jacovi, M., Maarek, Y.S., Pelleg, D., Shtalaim, M., Ur, S. (1998), The Shark-search algorithm. An application: tailored web site mapping, Computer 3etworks 30(1-7), pp 317-326

Hu, W., (2002), World Wide Web Search Technologies, Architectural Issues of Web-Enables Electronis

Business, edited by Shi Nansi for Idea Group Publishing Ifrim, G., Theobald, M., Weikum, G. (2005), Learning Word-to-Concept Mappings for Automatic Text

Classification, International Conference on Machine Learning

Jardine, N., Rijsbergen, C.J. (1971), The use of hierarchic clustering in information retrieval, Information Storage and Retrieval, 7(5), pp. 217-240

Joachims, T. (1998), Text Categorization with Support Vector Machines: Learning with Many Relevant

Features, Research Report of the unit no. VIII(AI), Computer Science Department of the University of Dortmund

Kandogan, E. (2001), Visualizing Multi-dimensional Clusters, Trends, and Outliers using Star

Coordinates, Proceedings of the KDD Conference, San Francisco, Califormia, USA Kahle, B. (1997), Preserving the internet, Scientific American. 276:3, p. 82-83

Kleinberg, J. (1998), Authoritative sources in a hyperlinked environment, Proceedings of the 9th ACM-

SIAM Symposium on Discrete Algorithms, pp 668-677 Koller, D., Sahami, M. (1996), Toward Optimal Feature Selection, Proceedings of the 13th International

Conference on Machine Learning, pp. 284-292, Morgan Kaufmann

Kosala, R., Blockeel, H. (2000), Web Mining Research: A Survey, SIGKDD Explorations, Vol. 2, No. 1, pp 1-13



Kumar, R., Raghavan, P., Rajagopalan, S., Sivakumar, D., Tomkins, A., Upfal, E. (2000), The Web as a graph, Proceedings of the 19th ACM SIGACT-SIGMOD-AIGART Symp. Principles of Database

Systems

Lafferty, J., McCallum, A., Pereira, F. (2001), Conditional random fields: Probabilistic models for segmenting and labeling sequence data , 18th International Conference on Machine Learning, 2001

Lawrence, S., Bollacker, K., Giles, C.L. (1999), Indexing and Retrieval of Scientific Literature,

Proceedings of the 8th International Conference on Information and Knowledge Management, pp 139-146

Lewandowski, D. (2005), Web searching, search engines and Information Retrieval. Information

Services and Use. 25:3-4/2005, p. 137-147 Li, X., Liu, B. (2003), Learning to classify text with positive and unlabelled data, Proceeding of IJCAI –

2003

Lim, L., Wang, M., Padmanabhan, S., Vitter, J.S., Agarwal, R. (2001), Characterizing Web Document Change, Lecture notes in Computer Science

Liu, R. L., Lin, W. J.(2005), Incremental mining of information interest for personalized web scanning,

Information Systems journal, 30(8), p. 630-648 Lu, S., Dong, M., Fotouhi, F. (2002), The semantic web: opportunities and challenges for next

generation web applications, Information Research, 7 (4)

Mitra, M., Singhal, A., Buckley, C. (1998), Improving automatic query expansion, Proceedings of the 21st ACM SIGIR Conference

Nelson, T. (1965), A file structure for the complex, the changing, and the indeterminate, ACM National

Conference, 84-100 Nicola, C., Gaussier, E., Goutte, C., Renders, J. M. (2003), “Word-Sequence Kernels”, Journal of

Machine Learning Research, Nº 3, pp 1053-1082

Nigam, K., McCallum, A.K., Thrun, S., Mitchell, T.M. (2000), Text classification from labeled and unlabeled documents using EM, Machine Learning, 39, pp 103-134

Olsen, K.A., Korfhage, R.R., Sochats, K.M., Spring, M.B., Williams, J.G. (1992), Visualization of a

Document Collection: The VIBE System, Information Processing & Management, Vol. 29, No. 1, pp 69-81

Orengo, V., Huyck, C. (2001), A Stemming Algorithm for the Portuguese Language, Proceedings of the

8th SPIRE O’Reily (2004), Web 2.0

Porter, M.F. (1980), “An algorithm for suffix stripping”, Program, 14, No. 3, pp 130-137

Richardson, M., Domingos, P. (2004) Combining Link and Content Information in Web Search, Washington University, Washington, USA

Rijsbergen, K. (1979), Information Retrieval, Butherworth

Sahami, M. (2004),The happy searcher: Challenges in the web information retrieval, Pacific Rim International Conference on Artificial Intelligence, 3157, p.3-12

Salton, G., Lesk, M.E. (1965), The SMART automatic document retrieval systems - an illustration,

Communications of the ACM, 8:6 (June 1965), p.391-398 Salton, G., McGill, M. (1983), Introduction to Modern Information Retrieval, McGraw-Hill

Salton, Wong, Yang (1975), A vector space model for automatic indexing. Communications of the ACM.

18:11 (1975). p. 613-620 Shen, G. (2005), Formal concepts and applications, PhD dissertation, Case Western Reserve University,

Ohio, USA

Shwarzkopf, E. (2003), Personalized Interaction with Semantic Information Portals, German Research Center for Artificial Intelligence

Siddiqui, Tanveer, J. (2006), Intelligent Techniques for Effective Information Retrieval (A Conceptual

Graph Based Approach), ACM SIGIR Forum. 40:2


Spangler, S., Kreulen, J.T., Lessler, J. (2003), Generating and Browsing Multiple Taxonomies Over a Document Collection, Journal of Management Information Systems, 19(4), p. 191-212

Viji, S. (2002), Term and Document Correlation and Visualization for a set of Documents, Technical

report, Stanford University Voorhees, E.M. (1998), Variations in Relevance Judgements and the Measurement of Retrieval

Effectiveness, Proceedings of the ACM SIGIR 1998 Conference

Wang, J., Lochovsky, F. (2003), “Web Search Engines”, Journal of ACM Computing Survey (accepted for revision)

Wolf, K.E. (1993), A First Course in Formal Concept Analysis, Advances in Statistical Software, 4, p.

429-438 Yang, Y. (1999), An Evaluation of Statistical Approaches t

o Text Categorization, Journal of Information Retrieval, vol. 1, nos. 1/2, pp 67-88

Yang, Y., Pederson, J. (1997), “A Comparative Study of Feature Selection in Text Categorization”, International Conference on Machine Learning

Yang, Y., Slattery, S., Ghani, R. (2002), A Study of Approaches to Hypertext Categorization, Kluwer

Academic Publishers, pp. 1-25 Zakos, J., Verma, B. (2006), A Novel Context-based Technique for Web Information Retrieval, World

Wide Web, 9(4), p. 485-503

Zamir, O., Etzioni, O. (1999), Grouper: A Dynamic clustering Interface to Web Search Results,

Proceedings of the 1999 World Wide Web Conference

Satisfying Information eeds on the Web: a Survey of Web ... · difficulties in the retrieval...

Documents

Transcript of Satisfying Information eeds on the Web: a Survey of Web ... · difficulties in the retrieval...